From: Hal Finney (hal@finney.org)
Date: Wed Feb 12 2003 - 17:34:53 MST
Wei mentioned Greg Egan's fictional Qusp, a computer for running an AI
in a deterministic way. I made some comments on this concept on the
everything-list, which you can read at
http://www.mail-archive.com/everything-list%40eskimo.com/msg04203.html.
I didn't think the idea made that much sense when I wrote that, but on
second thought I can see a possible reason for it.
The thing is, the Qusp is actually just an ordinary computer, as Wei
describes, something that does calculations in the same was as a Turing
machine or any other deterministic calculation model. All of our computers
today would be Qusps, at least if we had AI programs to run on them. So
what was there to design?
My new thought is to imagine that in this future time, all computers
are quantum computers. QCs are superior to classical computers in many
ways, but today we can't really build them. Let us suppose that in a
few decades the technical problems are solved. Then it is plausible
that this technology would be used for all high-performance computers.
In that case the inventor of the Qusp might have a point, in that AI
programs running on the computers of his day have an inherent quantum
randomness due to the fact that they are running on QCs. And likewise it
is plausible that our own brains are subject to quantum randomness due to
the messy and chaotic chemistry that is going on below the neural level.
Building a truly deterministic computer for the purpose of running an
AI would be a novel idea in that context.
So he wants to make a deterministic computer. Of course he can't just
dust off an old Pentium design from the year 2000! That would be as
slow as an abacus compared to the computers of his time. The challenge
is to come up with a new computer architecture that can match the
performance of the QCs in running an AI, without being subject to
the quantum non-determinism that is so essential to their operation.
Maybe he does still use quantum effects, but he has to do so cleverly
so that the high level brain functions are fully deterministic.
I think Robin Hanson wrote a paper some years ago about how QCs could
speed up AIs. They halve the depth of search trees, for example,
and searching is one of the key paradigms underlying many approaches
to AI. It could be far from trivial to achieve this kind of performance
improvement while maintaining determinism.
So I do think there is at least some justification for Egan imagining
a society where making a Qusp would be a project someone might try.
I still don't think that it really solves the problem of deciding things
both ways, or that that is truly a problem anyone should worry about.
But I acknowledge that some people do worry about it, so it does make
sense for Egan to write a story about the first guy who makes such a
machine, and the first AI who grows up with a Qusp as her brain.
Hal
This archive was generated by hypermail 2.1.5 : Wed Feb 12 2003 - 21:01:11 MST